Tidy Comfy UI workflow (analog photo)版本v1.0 (ID: 425263)

Clean, easy to look at text to image Comfy workflow. This setup works well for an analog photo style, especially with these LoRA:

Analog Style - v1.0 | Stable Diffusion LoRA | Civitai
Randommaxx Enhancer - v1.0 | Stable Diffusion LoRA | Civitai
DPO (Direct Preference Optimization) LoRA for XL and 1.5 - OpenRail++ - SDXL - V1.0 | Stable Diffusion LoRA | Civitai
Zest_XL - v1.0 | Stable Diffusion LyCORIS | Civitai
Midjourney mimic - v1.2 | Stable Diffusion LoRA | Civitai

Things are arranged mostly left to right, with a few things I consider in the Operations category broken out because you might want to adjust them more often.

I feel as though I get less prompt bleed by using the ComfyRoll Combine Prompt node, but if you want to swap it for another it's easy.

By default, sampler 2 has a very low denoise and is used more for detail/refining. It feeds the detailers and SD Upscale model. I find DemonCore or Juggernaut does a really nice job with this.

If you want to see what a bypassed node does, just highlight it, Ctrl+B, click the little button. They're all useful in ways but mostly I don't use them, including the detailers.

The filename will be printed with these data points: checkpoint 1, sampler 1, scheduler 1, LEB (swap this for your name I reckon) and the name of the last LoRA in the stacks, which is where I put one when testing it.

Other notes:

FreeU is very mild by default and fed to sampler 1. But I've been getting good results lately with these settings: b1: 1.0 b2: 1.2 s1: 0.6 s2: 0.4

LoRA stacks feed all samplers; this means SD upscale and detailers as well.

SD Upscale is set to 2x; I have tested loads of upscale models, and this (4x Superscale) is my favorite all-around. Some mess with the exposure, some ruin detail.

Show Text node is there if you want to copy out a version of the main prompt in one go; also very useful if you use wildcards to prompt.

ControlNet isn't great with SD XL, but I included it anyway. Second prompt is optional, but sometimes I want more chaos and a higher denoise for sampler 2. I feel like the ComfyRoll Combine Prompt node gives a little less prompt bleed and helps me keep the fields more organized. Finding all the models should be easy enough, I haven't renamed any. Only one I don't have a link handy for is Dice's Mega Detail.

You can turn down the steps on the second sampler and SD upscale if you're using a lower denoise; I'm using about 10 steps on sampler 2 with 0.2 denoise and 12 steps on the SD upscaler with 0.1 denoise. Can't see any difference from 15 steps, which I had used for both quite a lot.

描述:

训练词语: